\magnification=1200
\baselineskip=20pt
\nopagenumbers
\font\big=cmr12 scaled \magstep2

\centerline{\bf STANFORD UNIVERSITY}
\centerline{\bf DEPARTMENT OF STATISTICS}
\centerline{\big DEPARTMENTAL SEMINAR}
\bigskip

\baselineskip=12pt
\centerline{4:15 p.m., Thursday, June 29, 2000}
\centerline{Sequoia Hall Rm. 200}
\centerline{(Cookies at 3:45 in 1st Floor Lounge)}

\bigskip
\baselineskip=15pt
\centerline{\sl Peter Grunwald}
\centerline{\sl EURANDOM}
\centerline{\sl The Netherlands}
\bigskip
\centerline{\bf SAFE STATISTICS}
\bigskip


Statistical analysis of data more often than not results in inference of a
model that is ``wrong yet useful'': it is wrong in that it is really a gross
simplification of the process actually underlying the data; it is useful in
that decisions and predictions about future data taken on the basis of the
model are quite successful. Examples of this phenomenon abound: we assume
(conditional) independencies between random variables which really are
highly dependent (i.e. in speech recognition based on hidden Markov
models); we assume linear models for non-linear phenomena; we model noise
by a Gaussian also in cases where noise is really not Gaussian at all --
and we often get away with this.

Here we give a novel explanation of this phenomenon. We show that there
exist models (the simplest examples are the i.i.d exponential families)
that can be `safely' used for inference even if the `truth'
is not even close to any of the distributions under consideration. We
outline a general theory that allows us to determine under exactly what
circumstances it is useful or even advisable to use such overly
simple but `safe' models for the data at hand. We discuss relations to
other work on `robust' inference under misspecification.

\bye